| import pprint | |
| import pandas as pd | |
| from uniflow.flow.client import TransformClient | |
| from uniflow.flow.config import TransformOpenAIConfig | |
| from uniflow.op.prompt import Context | |
| from helpers import compare_strings_ignore_non_string | |
| from visualize_upload import visualize | |
| DEBUG = False | |
| def cluster(query, answers_data): | |
| answers = [] | |
| for answer in answers_data: | |
| answers.extend(answer["answer"]) | |
| data = [Context(context=query, excerpts=answers)] | |
| instruction = """ | |
| # Task: I am a researcher with a query about research papers. I have a list of excerpts from those papers. I need you to cluster each of these excerpts into a category based on the query. | |
| ## Input: | |
| 1. context: A brief query/context | |
| 2. excerpts: An list of excerpts from research papers. | |
| ## Evaluation Steps: | |
| ### Step 1 | |
| Go through each excerpt. For each excerpt, if there is an answer to the context/query that's not already captured by a category, create a category and add it to your category list. If the context has the word 'specific', make the category as specific as the excerpt. Repeat this process for each excerpt. The categories should be mutually exclusive. | |
| ### Step 2 | |
| Once you've gone through all the excerpts and you have a list of categories, go through the excerpts a second time, and this time assign each excerpt to a category. A single excerpt can be assigned to multiple categories. If there is no information relevant to any of the categories, please categorize the excerpt as "None". | |
| ## Response Format: Your response should only include two fields below: | |
| 1. categories: A list of all the generated categories. This is the output of Step 1 above. | |
| 2. clusters: An object, with each category as a key, and a list of all the excerpts as strings that fall into that category as the value. This is the output of Step 2 above. | |
| """ | |
| few_shot_examples = [ | |
| # Context( | |
| # context="Which types of batteries are discussed?", | |
| # excerpts=[ | |
| # "This investigation will shed lights on the tuneable chemical environments of transition-metal oxides for advanced cathode materials and promote the development of sodium-ion batteries.", | |
| # "Bi2Se3 was studied as a novel sodium-ion battery anode material because of its high theoretical capacity and high intrinsic conductivity.", | |
| # "Magnesium-ion batteries (MIBs) are considered strong candidates for next-generation energy-storage systems owing to their high theoretical capacity, divalent nature and the natural abundancy of magnesium (Mg) resources on Earth.", | |
| # "Magnesium-ion batteries (MIBs) have great potential in large-scale energy storage field with high capacity, excellent safety, and low cost.", | |
| # ], | |
| # categories=["Sodium-ion battery", "Magnesium-ion batteries"], | |
| # clusters={ | |
| # "Sodium-ion battery": [ | |
| # "This investigation will shed lights on the tuneable chemical environments of transition-metal oxides for advanced cathode materials and promote the development of sodium-ion batteries.", | |
| # "Bi2Se3 was studied as a novel sodium-ion battery anode material because of its high theoretical capacity and high intrinsic conductivity.", | |
| # ], | |
| # "Magnesium-ion batteries": [ | |
| # "Magnesium-ion batteries (MIBs) are considered strong candidates for next-generation energy-storage systems owing to their high theoretical capacity, divalent nature and the natural abundancy of magnesium (Mg) resources on Earth.", | |
| # "Magnesium-ion batteries (MIBs) have great potential in large-scale energy storage field with high capacity, excellent safety, and low cost.", | |
| # ], | |
| # }, | |
| # ), | |
| # Context( | |
| # context="Which 3D printing materials are discussed?", | |
| # excerpts=[ | |
| # "The current state of materials development, including metal alloys, polymer composites, ceramics and concrete, was presented", | |
| # "To this end, this work designs a novel 3D printing phase change aggregate to prepare concrete with prominent thermal capacity and ductility.", | |
| # "In this study, 15 commercial pure titanium samples are processed under different conditions, and the 3D pore structures are characterized by X-ray tomography", | |
| # "In this study, a support-less ceramic printing (SLCP) process using a hydrogel bath was developed to facilitate the manufacture of complex bone substitutes.", | |
| # ], | |
| # categories=[ | |
| # "metals", | |
| # "polymer composites", | |
| # "ceramics", | |
| # "concrete", | |
| # "phase change aggregate", | |
| # ], | |
| # clusters={ | |
| # "metals": [ | |
| # "The current state of materials development, including metal alloys, polymer composites, ceramics and concrete, was presented", | |
| # "In this study, 15 commercial pure titanium samples are processed under different conditions, and the 3D pore structures are characterized by X-ray tomography", | |
| # ], | |
| # "polymer composites": [ | |
| # "The current state of materials development, including metal alloys, polymer composites, ceramics and concrete, was presented" | |
| # ], | |
| # "ceramics": [ | |
| # "The current state of materials development, including metal alloys, polymer composites, ceramics and concrete, was presented", | |
| # "In this study, a support-less ceramic printing (SLCP) process using a hydrogel bath was developed to facilitate the manufacture of complex bone substitutes.", | |
| # ], | |
| # "concrete": [ | |
| # "The current state of materials development, including metal alloys, polymer composites, ceramics and concrete, was presented", | |
| # "To this end, this work designs a novel 3D printing phase change aggregate to prepare concrete with prominent thermal capacity and ductility.", | |
| # ], | |
| # "phase change aggregate": [ | |
| # "To this end, this work designs a novel 3D printing phase change aggregate to prepare concrete with prominent thermal capacity and ductility." | |
| # ], | |
| # }, | |
| # ), | |
| ] | |
| num_thread_batch_size = 1 | |
| config = TransformOpenAIConfig() | |
| config.prompt_template.instruction = instruction | |
| config.prompt_template.few_shot_prompt = few_shot_examples | |
| config.model_config.model_name = "gpt-4-1106-preview" | |
| config.model_config.response_format = {"type": "json_object"} | |
| config.model_config.num_call = 1 | |
| config.model_config.temperature = 0.0 | |
| config.model_config.num_thread = num_thread_batch_size | |
| config.model_config.batch_size = num_thread_batch_size | |
| cluster_client = TransformClient(config) | |
| output = cluster_client.run(data) | |
| if DEBUG: | |
| pprint.pprint(output) | |
| output_clusters = answers_data | |
| clusters = output[0]["output"][0]["response"][0]["clusters"] | |
| output_answer_category = [] | |
| for idx, paper in enumerate(answers_data): | |
| # Initialize an empty list to store the categories for each answer | |
| categories_per_answer = [] | |
| # Iterate over each answer | |
| for ans in paper["answer"]: | |
| categories = [] | |
| # Iterate over each category in clusters | |
| for category, texts in clusters.items(): | |
| # Check if the answer is in any of the texts related to the category | |
| if any(compare_strings_ignore_non_string(ans, text) for text in texts): | |
| if category not in categories_per_answer: | |
| categories.append(category) | |
| output_answer_category.append( | |
| {"paper": paper["paper"], "answer": ans, "category": category} | |
| ) | |
| if len(categories) == 0: | |
| categories.append("None") | |
| categories_per_answer.extend(categories) | |
| output_clusters[idx]["categories"] = categories_per_answer | |
| for output_cluster in output_clusters: | |
| if len(output_cluster["categories"]) == 0: | |
| output_cluster["categories"].append("None") | |
| df = create_category_df(output_clusters, answers_data) | |
| output_answer_category_df = pd.DataFrame(output_answer_category) | |
| visualize_output = visualize(output_clusters) | |
| return [output_clusters, df, visualize_output, output_answer_category_df] | |
| def create_category_df(cluster_output, answers_data): | |
| pd_data = { | |
| "Paper": [], | |
| "Excerpts": [], | |
| "Categories": [], | |
| } | |
| for i, paper in enumerate(cluster_output): | |
| pd_data["Paper"].append(paper["paper"]) | |
| pd_data["Excerpts"].append(", ".join(answers_data[i]["answer"])) | |
| pd_data["Categories"].append(", ".join(paper["categories"])) | |
| df = pd.DataFrame(pd_data) | |
| return df | |